|
|
|
|
@ -72,7 +72,9 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
|
|
|
|
|
use psb_base_mod
|
|
|
|
|
use amg_base_prec_type
|
|
|
|
|
use amg_c_inner_mod
|
|
|
|
|
|
|
|
|
|
#if defined(OPENMP)
|
|
|
|
|
use omp_lib
|
|
|
|
|
#endif
|
|
|
|
|
implicit none
|
|
|
|
|
|
|
|
|
|
! Arguments
|
|
|
|
|
@ -99,6 +101,9 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
|
|
|
|
|
integer(psb_ipk_) :: nrow, ncol, n_ne
|
|
|
|
|
integer(psb_lpk_) :: nrglob
|
|
|
|
|
character(len=20) :: name, ch_err
|
|
|
|
|
integer(psb_ipk_), save :: idx_soc1_p1=-1, idx_soc1_p2=-1, idx_soc1_p3=-1
|
|
|
|
|
integer(psb_ipk_), save :: idx_soc1_p0=-1
|
|
|
|
|
logical, parameter :: do_timings=.true.
|
|
|
|
|
|
|
|
|
|
info=psb_success_
|
|
|
|
|
name = 'amg_soc1_map_bld'
|
|
|
|
|
@ -114,6 +119,14 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
|
|
|
|
|
nrow = desc_a%get_local_rows()
|
|
|
|
|
ncol = desc_a%get_local_cols()
|
|
|
|
|
nrglob = desc_a%get_global_rows()
|
|
|
|
|
if ((do_timings).and.(idx_soc1_p0==-1)) &
|
|
|
|
|
& idx_soc1_p0 = psb_get_timer_idx("SOC1_MAP: phase0")
|
|
|
|
|
if ((do_timings).and.(idx_soc1_p1==-1)) &
|
|
|
|
|
& idx_soc1_p1 = psb_get_timer_idx("SOC1_MAP: phase1")
|
|
|
|
|
if ((do_timings).and.(idx_soc1_p2==-1)) &
|
|
|
|
|
& idx_soc1_p2 = psb_get_timer_idx("SOC1_MAP: phase2")
|
|
|
|
|
if ((do_timings).and.(idx_soc1_p3==-1)) &
|
|
|
|
|
& idx_soc1_p3 = psb_get_timer_idx("SOC1_MAP: phase3")
|
|
|
|
|
|
|
|
|
|
nr = a%get_nrows()
|
|
|
|
|
nc = a%get_ncols()
|
|
|
|
|
@ -133,33 +146,185 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
|
|
|
|
|
goto 9999
|
|
|
|
|
end if
|
|
|
|
|
|
|
|
|
|
if (do_timings) call psb_tic(idx_soc1_p0)
|
|
|
|
|
call a%cp_to(acsr)
|
|
|
|
|
if (do_timings) call psb_toc(idx_soc1_p0)
|
|
|
|
|
if (clean_zeros) call acsr%clean_zeros(info)
|
|
|
|
|
if (iorder == amg_aggr_ord_nat_) then
|
|
|
|
|
!$omp parallel do private(i)
|
|
|
|
|
do i=1, nr
|
|
|
|
|
ilaggr(i) = -(nr+1)
|
|
|
|
|
idxs(i) = i
|
|
|
|
|
end do
|
|
|
|
|
!$omp end parallel do
|
|
|
|
|
else
|
|
|
|
|
!$omp parallel do private(i)
|
|
|
|
|
do i=1, nr
|
|
|
|
|
ilaggr(i) = -(nr+1)
|
|
|
|
|
ideg(i) = acsr%irp(i+1) - acsr%irp(i)
|
|
|
|
|
end do
|
|
|
|
|
!$omp end parallel do
|
|
|
|
|
call psb_msort(ideg,ix=idxs,dir=psb_sort_down_)
|
|
|
|
|
end if
|
|
|
|
|
|
|
|
|
|
if (do_timings) call psb_tic(idx_soc1_p1)
|
|
|
|
|
|
|
|
|
|
!
|
|
|
|
|
! Phase one: Start with disjoint groups.
|
|
|
|
|
!
|
|
|
|
|
naggr = 0
|
|
|
|
|
icnt = 0
|
|
|
|
|
#if 0&&defined(OPENMP)
|
|
|
|
|
block
|
|
|
|
|
integer(psb_ipk_), allocatable :: bnds(:), locnaggr(:)
|
|
|
|
|
integer(psb_ipk_) :: myth,nths, kk
|
|
|
|
|
!$omp parallel shared(bnds,locnaggr,ilaggr,nr,naggr,diag,theta,nths) private(icol,val,myth,kk)
|
|
|
|
|
block
|
|
|
|
|
integer(psb_ipk_) :: ii,nlp,k,n,ia,isz, nc, i,j,m, nz, ilg, ip, rsz, minip
|
|
|
|
|
nths = omp_get_num_threads()
|
|
|
|
|
myth = omp_get_thread_num()
|
|
|
|
|
rsz = nr/nths
|
|
|
|
|
if (myth < mod(nr,nths)) rsz = rsz + 1
|
|
|
|
|
!!$ write(0,*) 'From thread : rsz ',myth,rsz
|
|
|
|
|
!$omp master
|
|
|
|
|
allocate(bnds(0:nths),locnaggr(0:nths))
|
|
|
|
|
locnaggr(:) = 0
|
|
|
|
|
bnds(0) = 1
|
|
|
|
|
!$omp end master
|
|
|
|
|
!$omp barrier
|
|
|
|
|
bnds(myth+1) = rsz
|
|
|
|
|
!$omp master
|
|
|
|
|
!!$ write(0,*) 'From master 1: ',bnds
|
|
|
|
|
do i=1,nths
|
|
|
|
|
bnds(i) = bnds(i) + bnds(i-1)
|
|
|
|
|
end do
|
|
|
|
|
!!$ write(0,*) 'From master 2: ',bnds
|
|
|
|
|
!$omp end master
|
|
|
|
|
!$omp barrier
|
|
|
|
|
|
|
|
|
|
!$omp do schedule(static)
|
|
|
|
|
do kk=0, nths-1
|
|
|
|
|
!!$ write(0,*) 'From thread ',myth,kk,bnds(kk),bnds(kk+1)-1
|
|
|
|
|
step1: do ii=bnds(kk), bnds(kk+1)-1
|
|
|
|
|
if (info /= 0) cycle
|
|
|
|
|
i = idxs(ii)
|
|
|
|
|
if ((i<1).or.(i>nr)) then
|
|
|
|
|
info=psb_err_internal_error_
|
|
|
|
|
call psb_errpush(info,name)
|
|
|
|
|
cycle step1
|
|
|
|
|
!goto 9999
|
|
|
|
|
end if
|
|
|
|
|
|
|
|
|
|
if (ilaggr(i) == -(nr+1)) then
|
|
|
|
|
nz = (acsr%irp(i+1)-acsr%irp(i))
|
|
|
|
|
if ((nz<0).or.(nz>size(icol))) then
|
|
|
|
|
info=psb_err_internal_error_
|
|
|
|
|
call psb_errpush(info,name)
|
|
|
|
|
cycle step1
|
|
|
|
|
!goto 9999
|
|
|
|
|
end if
|
|
|
|
|
|
|
|
|
|
icol(1:nz) = acsr%ja(acsr%irp(i):acsr%irp(i+1)-1)
|
|
|
|
|
val(1:nz) = acsr%val(acsr%irp(i):acsr%irp(i+1)-1)
|
|
|
|
|
|
|
|
|
|
!
|
|
|
|
|
! Build the set of all strongly coupled nodes
|
|
|
|
|
!
|
|
|
|
|
if (.false.) then
|
|
|
|
|
ip = 0
|
|
|
|
|
do k=1, nz
|
|
|
|
|
j = icol(k)
|
|
|
|
|
if ((bnds(myth)<=j).and.(j<=(bnds(myth+1)-1))) then
|
|
|
|
|
if (abs(val(k)) > theta*sqrt(abs(diag(i)*diag(j)))) then
|
|
|
|
|
ip = ip + 1
|
|
|
|
|
icol(ip) = icol(k)
|
|
|
|
|
end if
|
|
|
|
|
end if
|
|
|
|
|
enddo
|
|
|
|
|
|
|
|
|
|
!
|
|
|
|
|
! If the whole strongly coupled neighborhood of I is
|
|
|
|
|
! as yet unconnected, turn it into the next aggregate.
|
|
|
|
|
! Same if ip==0 (in which case, neighborhood only
|
|
|
|
|
! contains I even if it does not look like it from matrix)
|
|
|
|
|
!
|
|
|
|
|
disjoint = all(ilaggr(icol(1:ip)) == -(nr+1)).or.(ip==0)
|
|
|
|
|
if (disjoint) then
|
|
|
|
|
locnaggr(kk) = locnaggr(kk) + 1
|
|
|
|
|
do k=1, ip
|
|
|
|
|
ilaggr(icol(k)) = locnaggr(kk)
|
|
|
|
|
end do
|
|
|
|
|
ilaggr(i) = locnaggr(kk)
|
|
|
|
|
end if
|
|
|
|
|
else
|
|
|
|
|
ip = 0
|
|
|
|
|
minip = nr +1
|
|
|
|
|
do k=1, nz
|
|
|
|
|
j = icol(k)
|
|
|
|
|
if (abs(val(k)) > theta*sqrt(abs(diag(i)*diag(j)))) then
|
|
|
|
|
ip = ip + 1
|
|
|
|
|
icol(ip) = icol(k)
|
|
|
|
|
minip = min(icol(ip),minip)
|
|
|
|
|
end if
|
|
|
|
|
enddo
|
|
|
|
|
if (bnds(myth)<=minip) then
|
|
|
|
|
|
|
|
|
|
!
|
|
|
|
|
! If the whole strongly coupled neighborhood of I is
|
|
|
|
|
! as yet unconnected, turn it into the next aggregate.
|
|
|
|
|
! Same if ip==0 (in which case, neighborhood only
|
|
|
|
|
! contains I even if it does not look like it from matrix)
|
|
|
|
|
!
|
|
|
|
|
disjoint = all(ilaggr(icol(1:ip)) == -(nr+1)).or.(ip==0)
|
|
|
|
|
if (disjoint) then
|
|
|
|
|
!$omp critical(update_ilaggr)
|
|
|
|
|
disjoint = all(ilaggr(icol(1:ip)) == -(nr+1)).or.(ip==0)
|
|
|
|
|
if (disjoint) then
|
|
|
|
|
locnaggr(kk) = locnaggr(kk) + 1
|
|
|
|
|
do k=1, ip
|
|
|
|
|
ilaggr(icol(k)) = locnaggr(kk)
|
|
|
|
|
end do
|
|
|
|
|
ilaggr(i) = locnaggr(kk)
|
|
|
|
|
end if
|
|
|
|
|
!$omp end critical(update_ilaggr)
|
|
|
|
|
end if
|
|
|
|
|
endif
|
|
|
|
|
end if
|
|
|
|
|
end if
|
|
|
|
|
enddo step1
|
|
|
|
|
end do
|
|
|
|
|
!$omp end do
|
|
|
|
|
!$omp barrier
|
|
|
|
|
!$omp master
|
|
|
|
|
naggr = sum(locnaggr(0:nths-1))
|
|
|
|
|
!!$ write(0,*) 'NAGGR ',naggr, 'locnaggr ',locnaggr(0:nths-1)
|
|
|
|
|
do i=1,nths
|
|
|
|
|
locnaggr(i) = locnaggr(i) + locnaggr(i-1)
|
|
|
|
|
end do
|
|
|
|
|
do i=nths,1,-1
|
|
|
|
|
locnaggr(i) = locnaggr(i-1)
|
|
|
|
|
end do
|
|
|
|
|
locnaggr(0) = 0
|
|
|
|
|
!$omp end master
|
|
|
|
|
!$omp barrier
|
|
|
|
|
!$omp do schedule(static)
|
|
|
|
|
do kk=0, nths-1
|
|
|
|
|
do ii=bnds(kk), bnds(kk+1)-1
|
|
|
|
|
if (ilaggr(ii) > 0) ilaggr(ii) = ilaggr(ii) + locnaggr(kk)
|
|
|
|
|
end do
|
|
|
|
|
end do
|
|
|
|
|
!$omp end do
|
|
|
|
|
end block
|
|
|
|
|
!$omp end parallel
|
|
|
|
|
end block
|
|
|
|
|
!!$ write(0,*) 'Out of parallel looop NAGGR ',naggr
|
|
|
|
|
#else
|
|
|
|
|
step1: do ii=1, nr
|
|
|
|
|
if (info /= 0) cycle
|
|
|
|
|
i = idxs(ii)
|
|
|
|
|
if ((i<1).or.(i>nr)) then
|
|
|
|
|
info=psb_err_internal_error_
|
|
|
|
|
call psb_errpush(info,name)
|
|
|
|
|
goto 9999
|
|
|
|
|
cycle step1
|
|
|
|
|
!goto 9999
|
|
|
|
|
end if
|
|
|
|
|
|
|
|
|
|
if (ilaggr(i) == -(nr+1)) then
|
|
|
|
|
@ -167,7 +332,8 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
|
|
|
|
|
if ((nz<0).or.(nz>size(icol))) then
|
|
|
|
|
info=psb_err_internal_error_
|
|
|
|
|
call psb_errpush(info,name)
|
|
|
|
|
goto 9999
|
|
|
|
|
cycle step1
|
|
|
|
|
!goto 9999
|
|
|
|
|
end if
|
|
|
|
|
|
|
|
|
|
icol(1:nz) = acsr%ja(acsr%irp(i):acsr%irp(i+1)-1)
|
|
|
|
|
@ -195,7 +361,6 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
|
|
|
|
|
!
|
|
|
|
|
disjoint = all(ilaggr(icol(1:ip)) == -(nr+1)).or.(ip==0)
|
|
|
|
|
if (disjoint) then
|
|
|
|
|
icnt = icnt + 1
|
|
|
|
|
naggr = naggr + 1
|
|
|
|
|
do k=1, ip
|
|
|
|
|
ilaggr(icol(k)) = naggr
|
|
|
|
|
@ -204,16 +369,22 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
|
|
|
|
|
end if
|
|
|
|
|
endif
|
|
|
|
|
enddo step1
|
|
|
|
|
|
|
|
|
|
!!$ write(0,*) 'NAGGR ',naggr
|
|
|
|
|
#endif
|
|
|
|
|
if (debug_level >= psb_debug_outer_) then
|
|
|
|
|
write(debug_unit,*) me,' ',trim(name),&
|
|
|
|
|
& ' Check 1:',count(ilaggr == -(nr+1))
|
|
|
|
|
& ' Check 1:',naggr,count(ilaggr(1:nr) == -(nr+1)), count(ilaggr(1:nr)>0),&
|
|
|
|
|
& count(ilaggr(1:nr) == -(nr+1))+count(ilaggr(1:nr)>0),nr
|
|
|
|
|
end if
|
|
|
|
|
|
|
|
|
|
if (do_timings) call psb_toc(idx_soc1_p1)
|
|
|
|
|
if (do_timings) call psb_tic(idx_soc1_p2)
|
|
|
|
|
!
|
|
|
|
|
! Phase two: join the neighbours
|
|
|
|
|
!
|
|
|
|
|
! $ omp workshare
|
|
|
|
|
tmpaggr = ilaggr
|
|
|
|
|
! $ omp end workshare
|
|
|
|
|
! $ omp parallel do schedule(static) shared(tmpaggr,ilaggr,nr,naggr,diag,theta) private(ii,i,j,k,nz,icol,val,ip)
|
|
|
|
|
step2: do ii=1,nr
|
|
|
|
|
i = idxs(ii)
|
|
|
|
|
|
|
|
|
|
@ -244,8 +415,15 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
|
|
|
|
|
end if
|
|
|
|
|
end if
|
|
|
|
|
end do step2
|
|
|
|
|
! $ omp end parallel do
|
|
|
|
|
if (do_timings) call psb_toc(idx_soc1_p2)
|
|
|
|
|
if (debug_level >= psb_debug_outer_) then
|
|
|
|
|
write(debug_unit,*) me,' ',trim(name),&
|
|
|
|
|
& ' Check 1.5:',naggr,count(ilaggr(1:nr) == -(nr+1)), count(ilaggr(1:nr)>0),&
|
|
|
|
|
& count(ilaggr(1:nr) == -(nr+1))+count(ilaggr(1:nr)>0),nr
|
|
|
|
|
end if
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if (do_timings) call psb_tic(idx_soc1_p3)
|
|
|
|
|
!
|
|
|
|
|
! Phase three: sweep over leftovers, if any
|
|
|
|
|
!
|
|
|
|
|
@ -274,7 +452,6 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
|
|
|
|
|
end if
|
|
|
|
|
enddo
|
|
|
|
|
if (ip > 0) then
|
|
|
|
|
icnt = icnt + 1
|
|
|
|
|
naggr = naggr + 1
|
|
|
|
|
ilaggr(i) = naggr
|
|
|
|
|
do k=1, ip
|
|
|
|
|
@ -309,7 +486,7 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
|
|
|
|
|
endif
|
|
|
|
|
end if
|
|
|
|
|
end do
|
|
|
|
|
|
|
|
|
|
if (do_timings) call psb_toc(idx_soc1_p3)
|
|
|
|
|
if (naggr > ncol) then
|
|
|
|
|
!write(0,*) name,'Error : naggr > ncol',naggr,ncol
|
|
|
|
|
info=psb_err_internal_error_
|
|
|
|
|
@ -336,9 +513,14 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
|
|
|
|
|
nlaggr(:) = 0
|
|
|
|
|
nlaggr(me+1) = naggr
|
|
|
|
|
call psb_sum(ctxt,nlaggr(1:np))
|
|
|
|
|
if (debug_level >= psb_debug_outer_) then
|
|
|
|
|
write(debug_unit,*) me,' ',trim(name),&
|
|
|
|
|
& ' Check 2:',naggr,count(ilaggr(1:nr) == -(nr+1)), count(ilaggr(1:nr)>0),&
|
|
|
|
|
& count(ilaggr(1:nr) == -(nr+1))+count(ilaggr(1:nr)>0),nr
|
|
|
|
|
end if
|
|
|
|
|
!!$ write(0,*) nlaggr(1:np),'ILAGGR : ',ilaggr(1:nr)
|
|
|
|
|
|
|
|
|
|
call acsr%free()
|
|
|
|
|
|
|
|
|
|
call psb_erractionrestore(err_act)
|
|
|
|
|
return
|
|
|
|
|
|
sfilippone
1 year ago